Methodology, cited end-to-end.

LLM + VLM inference on Jetson / DRIVE

High-performance C++ runtime for LLM and VLM inference on resource-constrained NVIDIA platforms — FP8, NVFP4, INT4 quantization, EAGLE-3 speculative decoding, KV-cache compression. Demonstrated at CES 2026 with Bosch, ThunderSoft, MediaTek partner showcases.

Lightweight cross-platform inference

TensorFlow Lite's evolution into LiteRT — built-in quantization + compression, runs on Android, embedded Linux, microcontrollers via TFLite Micro. Default for cross-vendor mobile/edge deployments.

Cross-hardware AI inference

Cross-platform inference engine optimising AI models across CPUs, GPUs, NPUs, and specialised accelerators with minimal model modification. Deployed widely as a hardware-agnostic runtime.

PyTorch on microcontrollers + mobile

Bytecode VM with AOT compilation for PyTorch models — built for microcontrollers and embedded edge devices. Pairs with NVIDIA FLARE for federated fine-tuning on mobile.

Intel hardware-tuned inference

Optimised for Intel CPUs / GPUs / VPUs / FPGAs. Strong for industrial IoT vision (smart cameras, intelligent retail) and any deployment that's standardised on Intel silicon.

Production federated-learning runtime

Domain-agnostic SDK for federated learning. Hierarchical FL architecture for thousands of edge devices. Production deployments include Eli Lilly TuneLab, Taiwan MOHW national healthcare FL, and a Tri-Labs (Sandia / LANL / LLNL) federated AI pilot. Integrates with Flower; pairs with ExecuTorch for mobile FL.

Open FL framework + community

Cohesive approach to federated learning, analytics, and evaluation. Strong research community + extensive strategy library. Interoperates with FLARE so Flower-built apps run inside the FLARE runtime without modification.

Intel federated learning

Open-source FL implementation focused on sensitive-data deployments. Used in healthcare and regulated-industry pilots.

Kubernetes for edge devices

CNCF incubation project. Scales to 100,000 concurrent edge nodes managing 1,000,000+ active pods. The default when a customer's edge fleet is large enough to need cloud-native ops.

Edge-native K8s with offline ops

Brings edge computing capabilities to Kubernetes — edge nodes can run K8s without continuous cloud connectivity. Strong choice for intermittent / offline edge.

Lightweight Kubernetes

Optimised K8s with significantly reduced memory footprint — full K8s experience in resource-constrained environments. Lowest resource consumption among lightweight K8s distributions in 2026 benchmarks.

K8s leaf-device discovery + lifecycle

Built on Kubernetes Device Plugins; discovers small edge devices via ONVIF / udev / OPC UA handlers. Creates K8s services per device with HA when nodes lose network or fail.

Vendor-neutral edge IoT framework

LF Edge open-source edge platform. Modular reference services for device-data ingestion, normalisation, analysis, and sharing. The default integration substrate when a customer's deployment spans many vendor protocols.

Distributed application runtime

Portable runtime for distributed applications across cloud and edge. Built-in workflow, pub/sub, state, secrets, bindings, actors, distributed lock, cryptography. State of Dapr 2026 reports 20-40% developer productivity uplift.

Multi-protocol IoT messaging

Uniform API surface over MQTT, AMQP, HTTP, LoRaWAN — abstracts the protocol mess so application code can stay protocol-agnostic.

Standalone WebAssembly runtime

Bytecode Alliance reference runtime — leads in cold-start performance among JIT/AOT compilers (Jan 2026). 1-5 ms cold starts vs 100ms-1s+ for Linux containers — the 100x improvement that lets edge inference become serverless.

Edge-optimised WebAssembly runtime

Lightweight, high-performance WASM runtime for cloud-native, edge, and decentralised apps. Powers serverless apps, embedded functions, microservices, smart contracts, IoT devices. ~100x faster startup and ~1/100 the size of equivalent Linux containers.

Residential interoperability

Matter 1.4 (November 2024) added energy-management device classes. Thread provides the IPv6 mesh underneath. The default for Furcate's residential / SMB device class.

Long-range low-power IoT

Sub-GHz mesh for outdoor / industrial / agricultural deployments where Wi-Fi and cellular don't reach. Ultra-low-power, multi-km range, kilobit-class throughput.

Industrial-grade wireless

Private 5G slices + Time-Sensitive Networking (TSN) for deterministic latency on the factory floor. The wireless backbone Furcate's industrial customers run on.

Hardware root of trust

Trusted Platform Module 2.0 + Trusted Execution Environment (Intel SGX, ARM TrustZone) for secure boot, attested device identity, and confidential inference. The hardware backstop for Furcate's sovereign-by-design posture.

Robot middleware

Robot Operating System 2 — DDS-based publish/subscribe middleware for cobots, AMRs, manipulators, drones. Native integration so Industry's robotic workflows speak the same protocol as the rest of the fleet.

Distributed model serving

Triton for general inference serving, vLLM for LLM throughput, Ray Serve for distributed Python services. The serving substrate behind Furcate's edge gateways.